CN111062466B - Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network - Google Patents
Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network Download PDFInfo
- Publication number
- CN111062466B CN111062466B CN201911263143.0A CN201911263143A CN111062466B CN 111062466 B CN111062466 B CN 111062466B CN 201911263143 A CN201911263143 A CN 201911263143A CN 111062466 B CN111062466 B CN 111062466B
- Authority
- CN
- China
- Prior art keywords
- antenna
- data
- grid
- neural network
- grids
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R29/00—Arrangements for measuring or indicating electric quantities not covered by groups G01R19/00 - G01R27/00
- G01R29/08—Measuring electromagnetic field characteristics
- G01R29/0807—Measuring electromagnetic field characteristics characterised by the application
- G01R29/0814—Field measurements related to measuring influence on or from apparatus, components or humans, e.g. in ESD, EMI, EMC, EMP testing, measuring radiation leakage; detecting presence of micro- or radiowave emitters; dosimetry; testing shielding; measurements related to lightning
- G01R29/0821—Field measurements related to measuring influence on or from apparatus, components or humans, e.g. in ESD, EMI, EMC, EMP testing, measuring radiation leakage; detecting presence of micro- or radiowave emitters; dosimetry; testing shielding; measurements related to lightning rooms and test sites therefor, e.g. anechoic chambers, open field sites or TEM cells
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Electromagnetism (AREA)
- Mobile Radio Communication Systems (AREA)
- Variable-Direction Aerials And Aerial Arrays (AREA)
Abstract
The invention relates to a prediction method for cell field intensity distribution after antenna adjustment based on parameters and BP neural network, which comprises the following steps: (1) Collecting user MDT data, and reporting the user MDT data by User Equipment (UE); (2) Processing the MDT data collected in the step (1) and removing abnormal values; (3) Dividing a cell into square grids, dividing the divided square grids into two types, wherein the first type of grids are grids with sufficient RSRP data and are used for training a neural network; the second type of grids are grids with insufficient RSRP data; (4) Training by adopting a three-layer neural network, and outputting insufficient grid power of RSRP data; (5) And obtaining the average received power of each grid after the antenna azimuth rotation according to the formula of the average received power of each grid after the antenna azimuth rotation, namely, predicting the field intensity distribution of the cell after the antenna adjustment. The method has high accuracy and small error.
Description
Technical Field
The invention relates to an electromagnetic wave field intensity prediction technology, in particular to a prediction method for cell field intensity distribution after antenna adjustment based on parameters and a BP neural network.
Background
The research of the field intensity prediction model mainly aims at meeting the scientific planning of the mobile communication network, reasonable site selection can effectively provide coverage, eliminate coverage blind areas and provide convenience for future network optimization.
The prediction research work of electric wave propagation at home and abroad is widely developed from seventies, a plurality of analysis models are established by fumbling, different propagation models are needed under different scenes, a Road model is developed successively, an Oknmura-Hata model, an ECAC model, a Palmer model, a Murphy model suitable for suburban scenes, a cell ray tracing model and the like are adopted. More common, precisely predicted field strengths are deterministic models, i.e., propagation environments, building information, etc., as accurate input information, i.e., ray Tracing (RT) models. RT transport uses Fresnel theory, reflectance and geometry diffraction theory (Geometrical Theory of Diffraction, GTD) and consistency diffraction theory (Uniform Theory of Diffraction, UTD). Because of the large amount of accurate environmental data required, ray tracing models are generally limited to indoor and are not widely used in outdoor scenes.
The Artificial Neural Network (ANN) can exactly make up for the defect, has the advantage of being applied to field intensity prediction, can obtain more accurate field intensity values by utilizing a plurality of related parameters which are easy to obtain under the condition that the shape, the parameters and the structural characteristics of a building are unknown, and can flexibly adapt to different scenes. The principle is that available parameters are used as the input of the neural network, different neurons are arranged on each layer through different layers, a specific transfer function is used for reaching the output layer of the neural network, and then the actual measurement data is used for training, namely, the parameters of the neural network are adjusted to find the optimal transfer function. ANNs comprise a series of interconnected elementary units, called neurons or nodes.
At present, a field intensity prediction method of an outdoor micro cell is introduced in related documents, the topographic information parameters of the cell are input into a neural network for training, and the accuracy of a model is verified by using measured data. The literature proposes an outdoor hybrid prediction model that combines an empirical model with an ANN model, taking into account Line of Sight (LOS) and non-Line of Sight (Non Line of Sight, NLOS) situations at the transmitting and receiving end, respectively, such models being generally used for different terrain types, including urban, suburban, etc. The related literature describes an indoor prediction method that compares the effects of multiple types of ANNs for field strength prediction. The ANN method is also adopted in the related literature, and machine learning is used for classifying different terrains, so that necessary scene information can be automatically extracted for training. The related literature describes a hybrid differential prediction model that uses a coarse scene model and a small amount of accurate prediction values to train a multi-layer perceptron neural network so as to obtain accurate field strength prediction effects with low computational complexity. The various algorithms described above can naturally achieve higher accuracy, however, subject to the support of accurate environmental information, and some algorithms do not have true data to verify.
Therefore, in order to address the above-mentioned shortcomings, it is necessary to develop a prediction method for the distribution of the cell field intensity after the antenna adjustment based on parameters and the BP neural network, and utilize a large amount of user data before the antenna adjustment to predict the received power of any geographic position in the effective coverage area of the antenna after the antenna azimuth adjustment through ANN, without depending on environmental data, and the result of comparing with the measured data indicates that the prediction error is within a reasonable accuracy range.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a prediction method for the distribution of the cell field intensity after the antenna is adjusted based on parameters and BP neural network, the received power of any geographic position in the effective coverage area of the antenna is predicted by ANN, and the received power of the user in the effective coverage area after the antenna azimuth angle is adjusted is predicted based on an electric wave propagation formula and antenna pattern gain change, and environmental data is not needed to be relied on; and the prediction error is small.
In order to solve the technical problems, the invention adopts the following technical scheme: the method for predicting the cell field intensity distribution after the antenna adjustment based on the parameters and the BP neural network comprises the following steps:
(1) Collecting user MDT data, and reporting the user MDT data by User Equipment (UE);
(2) Processing the MDT data collected in the step (1) and removing abnormal values;
(3) Dividing a cell into square grids, dividing the divided square grids into two types, wherein the first type of grids are grids with sufficient RSRP data and are used for training a neural network; the second type of grids are grids with insufficient RSRP data;
(4) Training by adopting a three-layer neural network, and outputting insufficient grid power of RSRP data;
(5) And obtaining the average received power of each grid after the antenna azimuth rotation according to the formula of the average received power of each grid after the antenna azimuth rotation, namely, predicting the field intensity distribution of the cell after the antenna adjustment.
With the above-mentioned technical solution, since the conventional radio wave field intensity prediction model, that is, the ray tracing model (RT) model, requires a large amount of accurate environmental data, and its application range is limited only to indoor and outdoor applications have not been widely used. The Artificial Neural Network (ANN) can exactly make up the defects and limitations of ray tracing, can obtain relatively accurate field intensity values by utilizing a plurality of relevant parameters which are easy to obtain under the condition that the shape, the parameters and the structural characteristics of a building are unknown, and can flexibly adapt to different scenes. The technical scheme provides a prediction method based on parameter estimation and Back Propagation (BP) neural network without inputting scene information influencing electromagnetic radiation; dividing a cell into a plurality of grids, regarding the received power of User Equipment (UE) in each grid as a random variable meeting quasi-normal distribution, reading reference signal received power (Reference Signal Receiving Power, RSRP) data reported by all the UE in the grid, properly screening and calculating the mean value and standard deviation of the reference signal received power (Reference Signal Receiving Power, RSRP) data, respectively serving as parameter estimation values of the quasi-normal distribution, training a BP neural network by using the RSRP mean value of the UE in all the grids, and obtaining the received power of each grid after antenna adjustment based on a radio wave propagation formula and antenna pattern gain variationThe rate average value is the predicted value of the average value of the UE received power in the grid, and the predicted value of the standard deviation is regarded as the same as that before the antenna adjustment; using the available parameters as the input of the neural network, setting different neurons through different layers, reaching the output layer of the neural network through a specific transfer function, and training by using measured data; the prediction error is within a reasonable precision range without depending on environmental data and comparing the prediction error with actual measurement data; taking the coordinates of the receiving points and MDT data as the input of the ANN and performing data training, wherein the final output of the ANN is the predicted value of the RSRP of the User Equipment (UE); the gradient of the cost function of the ANN is rapidly calculated by adopting back propagation, and the input weight value of each layer of neural network in the ANN is automatically corrected according to the gradient, so that the prediction accuracy of the ANN is improved; the invention uses the minimization drive tests (Minimization Drive Test, MDT) data collected by different base station cells to carry out parameter estimation, takes the geographic coordinates of the user points as the input parameters of the neural network, trains and outputs the received power (Reference Signal Receiving Power, RSRP) value of the reference signals of the user points through the BP neural network, and predicts the received power of the users in the effective coverage area after the antenna azimuth angle adjustment based on the electric wave propagation formula and the gain change of the antenna pattern; the Back Propagation (BP) neural network belongs to a feedforward neural network and has the characteristics of high learning rate and strong self-adaption capability. The device consists of an input layer, an intermediate layer and an output layer, wherein the intermediate layer can be a plurality of layers. The specific learning method is that when a pair of learning models are provided for a network, the input of the network reaches a neuron to generate a connection Weight (Weight), then the connection Weight is corrected layer by layer from an output layer to an intermediate layer, the process is repeatedly performed until the error between the ANN output and the target output falls below a set threshold value, or the training frequency is higher than a preset upper limit, and the learning process can be considered to be completed. A great advantage of BP neural networks is the strong nonlinear mapping capability, which, theoretically, can approximate a nonlinear function with arbitrary accuracy as long as the number of neurons in the hidden layer is sufficiently large for a BP network of three or more layers. When the BP neural network selects hidden layer neurons, attention should be paid to selecting proper number of neurons, and too few neurons will cause the networkToo much tends to result in network overfitting, i.e., over fitting. Formula (VI)Gives the basic logical structure of ANN, where fi is a transfer function, usually a nonlinear function, usually Heaviside, sigmoid or Gaussian function, etc.; yi represents the output of the output node i; xj represents the input of the input node i; wij represents the link weights of nodes i and j; θi represents the offset of node i. The three-layer ANN is adopted, and the activation function of the second-layer neural network is as follows: h θ ＝tan(sigmoid(θ T ·[1x m y m ] T ) A) is provided; the activation function of the third layer neural network adopts the following steps: />The method comprises the steps of rasterizing an area covered by an antenna, carrying out parameter estimation on a large amount of user data reported before antenna azimuth angle adjustment in each grid, obtaining characteristics of the user data, screening out reasonable training data through corresponding data processing, predicting the received power of any geographic position in the effective coverage area of the antenna through an ANN, and predicting the received power of the user in the effective coverage area after antenna azimuth angle adjustment based on an electric wave propagation formula and antenna pattern gain change without depending on environmental data. The actual measurement results of two long term evolution (Long Term Evolution, LTE) base stations of a certain operator are called to evaluate the performance of the proposed prediction algorithm, and the result shows that the prediction average error of the algorithm is 5-8 dB, and the width of a 99% confidence interval is about 15dB.
As a preferable technical solution of the present invention, the user MDT data in the step (1) includes longitude θ and latitude of the userTime advance T a And reference signal received strength RSRP.
As a preferred technical scheme of the present invention, the step (2) specifically includes the following steps:
s21, taking the point with the altitude less than zero as abnormal data, and deleting the abnormal data;
s22, setting the altitude data as a constant 0, carrying out two-dimensional ink card support projection on the longitude and latitude of the reported user data, and enabling the north direction to correspond to the x-axis direction of the coordinates, wherein the azimuth angle of the antenna is the anticlockwise included angle between the vector from the antenna to the test point and the positive axis of the x; subtracting the antenna coordinates from the user point coordinates to obtain the relative coordinates of the user points with the antenna as an origin as new coordinates;
s23, dividing an antenna radiation coverage area into a plurality of grids, wherein the size of the grids is in the same order of magnitude as the precision of a Global Positioning System (GPS); for the RSRP value reported by different User Equipments (UE) in each grid, taking the RSRP values with the mean value mu and standard deviation sigma respectively set as the RSRP values of the User Equipments (UE) at the position as the mean value and standard deviation of random variables, whereinm is the total number of user equipments reporting RSRP values within one grid,and the average value mu and the standard deviation sigma of the data reported by the users in the grid are calculated, and the data positioned in [ mu-3 sigma, mu+3 sigma is obtained]Data within the range, data outside the range are used as outliers and deleted. The altitude information uploaded by the UE is usually low in precision or is not uploaded, so that the altitude information is set to be constant 0 regardless of the altitude; therefore, the longitude and latitude of the reported user data are subjected to two-dimensional ink card support projection, the north direction corresponds to the coordinate x positive axis direction, the azimuth angle, that is, the anticlockwise included angle between the vector from the antenna to the test point and the x positive axis, for example, the azimuth angle 120 degrees indicates that the projection of the UE position on the xoy plane is located in the direction anticlockwise rotated 120 degrees from the x positive axis direction.
As a preferred technical scheme of the present invention, the step (3) specifically includes the following steps:
dividing a cell into square grids of n M multiplied by n M, replacing the coordinates of User Equipment (UE) positioned in a certain grid with the coordinates of the center of the grid, taking the received power of the User Equipment (UE) in the grid as a random variable, and calculating the mean value and standard deviation of the random variable based on RSRP data reported by all the User Equipment (UE) in the grid; dividing all grids into two types according to different RSRP data amounts, wherein the first type of grids are grids with sufficient RSRP data and are used for training a neural network; the second type of grids are grids with insufficient RSRP data and even unreported RSRP data, and the corresponding grid power is predicted through a neural network.
As a preferred technical scheme of the present invention, the step (4) specifically includes the following steps:
training a grid with sufficient RSRP data by using a three-layer neural network, wherein the input layer is the coordinates (x a ,y a ) Wherein a represents a grid index number; the number of neurons contained in the middle hidden layer and the scale of a sample set adopted by training are in positive correlation, the output layer is a predicted value of the average value of the received power of the second type of grids, and constant term neurons of the input layer and the hidden layer belong to a bias unit;
wherein the activation function employed by each neuron of the hidden layer is:
wherein θ is hidden The weight vector from the input layer to the hidden layer neuron forms a weight matrix from the input layer to the hidden layer; the activation function complicates the input features to obtain more complex features, thereby fitting out more complex fitting functions from input to prediction output to improve the accuracy of the prediction result;
the activation function adopted by the output layer is a linear function:
where x is the vector of the output components of the hidden layer, θ output From hidden layer to output layer neuronsWeight vectors, weight vectors of neurons of all output layers form a weight matrix from a hidden layer to the output layer; different weight matrixes are arranged between layers of the neural network, the values of the weight matrixes are initialized randomly, the cost function is continuously optimized along with training of the neural network to obtain a final optimal weight matrix value, and finally, the predicted value of the average value of all grid received power in an effective radiation area of the antenna before the antenna is adjusted is obtained.
As a preferred technical scheme of the present invention, the step (5) specifically includes the following steps:
obtaining an average received power formula of each grid after antenna azimuth rotation by adopting an electric wave propagation formula through an electric wave propagation model in a field intensity prediction model, and predicting user received power in an effective coverage area after antenna azimuth adjustment based on the electric wave propagation formula and antenna pattern gain change, namely obtaining prediction of cell field intensity distribution after antenna adjustment;
the electric wave propagation formula is as follows:
pt in this formula is the antenna transmit power,GT is antenna gain, lambda is wave wavelength, d is free space distance from antenna to user point, alpha is path loss index, pr is receiving power of receiving point;
the average received power per grid after antenna azimuth rotation is given by:
wherein the method comprises the steps ofAnd->Respectively representing the rotation front and back (azimuth angle, downtilt angle) of the antenna, comparing the rotation front and back (azimuth angle, downtilt angle) with the average value of the user measured data reported after the antenna posture adjustment, and taking the obtained error as the error of the average value of the receiving power predicted by the algorithm. Wherein->Is used for calculating a path loss index alpha from an antenna to a user point (UE) after a neural network outputs a predicted value Pt; calculating the RSRP value of the original user point (UE) after the antenna azimuth angle is adjusted by using the calculated alpha; meanwhile, as long as the antenna attitude adjustment amplitude is not very large, the standard deviation of the received power is basically not influenced, and based on the consideration, the standard deviation before adjustment is directly used as a predicted value of the standard deviation after adjustment.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1) No scene information affecting electromagnetic radiation, such as building shape, geometry, dielectric constant, conductivity, etc., is input;
2) The probability statistical model is adopted, so that the characteristics of the RSRP sample data are more consistent with those of the actually collected RSRP sample data, and the RSRP sample data are necessarily influenced by the inconsistency of the performance, the position, the gesture and the channel fading of the UE and the random error of GPS positioning.
3) For areas where no UE arrives, i.e. areas where no RSRP data is reported, can also be predicted.
Drawings
The technical scheme of the invention is further described below with reference to the accompanying drawings:
fig. 1 is an antenna pattern provided by an operator, wherein fig. (a) is an H-plane pattern and fig. (b) is a V-plane pattern;
FIG. 2 is an antenna pattern using two-dimensional rose surface fitting;
fig. 3 is a histogram of base station user received power distribution;
fig. 4 is a diagram of a neural network architecture used to predict UE received power for a signal coverage area after antenna pose adjustment based on parameter estimation and the neural network;
FIG. 5 is a simplified algorithm flow chart for predicting UE received power for a signal coverage area after antenna pose adjustment based on parameter estimation and neural network;
fig. 6 is an error distribution diagram of the difference between the received power of the UE in the domain predicted by the algorithm of the present invention and the RSRP reported by the UE in the signal coverage area as the prediction error, for a certain base station of the operator of embodiment 1, where the antenna azimuth angle is adjusted to 150 degrees;
fig. 7 is a statistical histogram of RSRP values reported by UEs within the antenna azimuth angle adjusted to 150 degrees for a certain base station of the operator of embodiment 1;
fig. 8 is a statistical histogram of intra-domain UE received power predicted values obtained by using the algorithm of the present invention for a base station of the operator of example 1, which adjusts the antenna azimuth angle to 150 degrees;
fig. 9 is an error distribution diagram of the difference between the UE received power in the domain predicted by the algorithm of the present invention and the RSRP reported by the UE in the signal coverage area as a prediction error, for a certain base station of the operator of embodiment 2, where the antenna azimuth angle is adjusted to 150 degrees;
fig. 10 is a statistical histogram of RSRP values reported by UEs in the field of a base station of the operator of embodiment 2, wherein the antenna azimuth is adjusted to 150 degrees;
fig. 11 is a statistical histogram of intra-domain UE received power predicted values obtained by using the algorithm of the invention for a certain base station of the operator of example 2, adjusting the antenna azimuth angle to 150 degrees;
fig. 12 is a diagram showing RSRP value distribution reported by an intra-domain UE when an antenna azimuth angle is adjusted to 150 degrees for a certain base station of the operator in embodiment 2;
fig. 13 is a diagram showing predicted values of UE received power in the domain obtained by the algorithm of the present invention, which is obtained by adjusting the antenna azimuth angle to 150 degrees for a certain base station of the operator of example 2. The two graphs of FIG. 12 and FIG. 13 reflect intuitively that the values and the distribution trends of the two graphs are basically consistent;
fig. 14 is a graph of RSRP statistics and normalization reported by a UE in a certain grid when the antenna azimuth angle is unchanged for a certain base station of the operator of embodiment 1;
fig. 15 is a graph of RSRP statistics and normalization reported by a UE in a certain grid when the antenna azimuth angle is unchanged for a certain base station of the operator of embodiment 2;
fig. 16 is a statistical histogram of RSRP reported by a UE in a certain grid before antenna adjustment for a certain base station of the operator of embodiment 2;
fig. 17 is a statistical histogram of RSRP reported by the UE in the grid after antenna adjustment for a certain base station of the operator of example 2.
Detailed Description
The present invention will be further described in detail with reference to the drawings and examples, which are only for the purpose of illustrating the invention and are not to be construed as limiting the scope of the invention.
In order to verify the correctness of the algorithm and evaluate the accuracy of the algorithm, two base stations are taken as an example, MATLAB programming is adopted, the field intensity distribution of the coverage area of the antenna after the antenna posture is adjusted is predicted, and a reference is provided for wireless network planning. Two embodiments are adopted, one base station of an operator has fewer user points, distributed and scattered, and the scene is simple; and in the other case, a certain base station of a certain operator has more user points, and the distribution is mainly concentrated in a street cell, so that the scene is complex. Here, the altitude is not considered, only the horizontal plane coordinates are considered, and only for the case where the antenna azimuth adjustment amplitude is not large, the amount of change in the path loss index α can be considered negligible based on the principle of multipath fading.
An antenna is a radiation source of electric waves, and the directivity of the antenna reflects the distribution of the ability to radiate (or receive) electromagnetic waves in various directions, and the characteristic curve of the directivity of the antenna is generally represented by a pattern.
Because the general trend of the antenna pattern is very similar to that of the rose curve, the two-dimensional rose curve fitting method is adopted to fit the antenna pattern, and different characteristic rose curves are formed by adopting different rose curve parameters, so that the method can be used for fitting the antenna pattern with different characteristics; antenna vendor proposal based on certain operatorThe antenna pattern is sampled, then the two-dimensional rose surface fitting method is adopted to fit the antenna pattern, and the optimal parameters of the fitting function are found based on the least square method, so that more accurate prediction is made on the antenna pattern function, and gain values in any direction are obtained, namely, in an electric wave propagation formula
Example 1: the method for predicting the cell field intensity distribution after the antenna adjustment based on the parameters and the BP neural network is shown in a figure 1, wherein the figure is an antenna pattern provided by an antenna supplier, and the figure (a) is an H-plane pattern and the figure (b) is a V-plane pattern; FIG. 2 is a diagram showing a pattern obtained by two-dimensional rose surface fitting; according to the User data reported by the User Equipment (UE), the frequency distribution of the received power in the whole cell is counted, and fig. 3 is a distribution diagram of RSRP (dBm) reported by a certain base station of a certain operator in a certain day; the received power distribution of the UE in the whole cell can be reflected by fig. 3, which shows irregular random distribution, and the probability distribution is difficult to quantitatively analyze. In this case, we will use a method of dividing the grid to perform a finer analysis on the UE power distribution in the grid and combine the neural network to predict the average RSRP value of the grid; the neural network architecture adopted is shown in fig. 4, and the simplified algorithm flow chart is shown in fig. 5;
the method specifically comprises the following steps:
(1) According to the industrial parameter record, the maximum transmitting power of the base station (base station 1) is 52dBm, the channel number of the central carrier frequency is 38496, the antenna gain is 14dBi, 8756 user points are totally used when the initial azimuth angle of the antenna is 130 degrees, the field intensity distribution of the coverage area of the antenna needs to be predicted if the azimuth angle is adjusted to 150 degrees, and 4803 user points can be used for comparing and calculating the prediction error when the azimuth angle of the antenna is 150 degrees. The method is characterized in that based on the relation [15] between the Poyting vector and the electromagnetic field intensity, the RSRP decibel value is indirectly adopted for prediction;
(2) Processing the MDT data collected in the step (1) and removing abnormal values;
s21, regarding the point with the altitude less than zero as abnormal data, and deleting the abnormal data;
s22, because the altitude information uploaded by the UE is usually low in precision or is not uploaded with such data, the altitude is not considered, and the altitude information is set to be constant 0, so that the longitude and latitude of the reported user data are subjected to two-dimensional ink card support projection, the north direction corresponds to the coordinate x positive axis direction, the azimuth angle, namely the anticlockwise included angle between the vector from the antenna to the test point and the x positive axis, for example, the azimuth angle 120 degrees indicates that the projection of the position of the UE on the xoy plane is positioned in the anticlockwise rotation direction 120 degrees from the x positive axis direction; finally subtracting the antenna coordinates from the user point coordinates to obtain the relative coordinates of the user points with the antenna as the origin as new coordinates;
s23, for the RSRP values reported by different UE in each grid, taking the mean value mu and the standard deviation sigma of the RSRP values as the mean value and the standard deviation of the RSRP values as random variables of the UE at the position, wherein m is the total number of user equipments reporting RSRP value in one grid, +.> And the average value mu and the standard deviation sigma of the data reported by the users in the grid are calculated and are positioned in [ mu-3 sigma, mu+3 sigma ]]Data should occur with a high probability beyond which it is considered outliers and these outliers are deleted;
(3) Performing n M × n M grid division on user areas before and after adjustment of the antenna azimuth angle of the base station 1, and performing data processing on each grid user point data;
for the coordinates of the UE of each grid, the coordinates of the center of the grid (x m ,y m ) Instead, the received power of the UE in the grid is regarded as a random variable, and the average of the random variable is estimated based on RSRP data reported by all the UEs in the gridValues and standard deviations. Dividing all grids into two types according to different RSRP data amounts, wherein the first type is a grid with sufficient RSRP data, and forming a training data set (x) by adopting center coordinates of the grid and average values of RSRP values reported by all UE in the grid m ,y m, RSRP ave ) Wherein m is a grid index number, RSRP ave The average RSRP value of the m-number grids is used for training a neural network, the second class is grids with insufficient RSRP data and even unreported RSRP data, and the second class predicts the corresponding grid power of the second class through the neural network;
(4) Training the designed neural network by using a training data set to obtain an optimal weight value corresponding to each layer of the neural network;
training a grid with sufficient RSRP data by using a three-layer neural network, wherein the input layer is the coordinates (x a ,y a ) Wherein a represents a grid index number; the number of neurons contained in the middle hidden layer and the scale of a sample set adopted by training are in positive correlation, the output layer is a predicted value of the average value of the received power of the second type of grids, and constant term neurons of the input layer and the hidden layer belong to a bias unit;
wherein the activation function employed by each neuron of the hidden layer is:
wherein θ is hidden The weight vector from the input layer to the hidden layer neuron forms a weight matrix from the input layer to the hidden layer; the activation function complicates the input features to obtain more complex features, thereby fitting out more complex fitting functions from input to prediction output to improve the accuracy of the prediction result;
the activation function adopted by the output layer is a linear function:
where x is the vector of the output components of the hidden layer, θ output The weight vectors of the neurons of all the output layers form a weight matrix from the hidden layer to the output layer; different weight matrixes are arranged between layers of the neural network, the values of the weight matrixes are initialized randomly, a final weight matrix value is obtained along with training of the neural network, and finally, the predicted value of the average value of all grid received power in an effective radiation area of the antenna before adjustment is obtained;
(5) Obtaining an average received power formula of each grid after antenna azimuth rotation by adopting an electric wave propagation formula through an electric wave propagation model in a field intensity prediction model, and predicting user received power in an effective coverage area after antenna azimuth adjustment based on the electric wave propagation formula and antenna pattern gain change, namely obtaining prediction of cell field intensity distribution after antenna adjustment;
the electric wave propagation formula is as follows:
pt in this formula is the antenna transmit power,GT is antenna gain, lambda is wave wavelength, d is free space distance from antenna to user point, alpha is path loss index, pr is receiving power of receiving point;
the average received power per grid after antenna azimuth rotation is given by:
wherein the method comprises the steps ofAnd->Respectively representing the rotation front and back (azimuth angle, downward inclination angle) of the antenna, comparing the rotation front and back (azimuth angle, downward inclination angle) with the average value of the user measured data reported after the antenna posture adjustment, and taking the obtained error as the error of the average value of the receiving power predicted by the algorithm; i.e. the algorithm accuracy measure.
The prediction result shows that the prediction error of the distribution area with the average value of the absolute value of the error reaching 5.00dB and more than 96% is less than 15dB by adopting the algorithm to predict the field intensity distribution after the base station antenna is adjusted. The difference between the received power of the UE predicted by the algorithm and the RSRP reported by the UE in the signal coverage area is used as a prediction error, the distribution of the difference is shown in fig. 6, the distribution is reflected in the diagram, and the prediction error of most areas is smaller.
Wherein the prediction error for most of the regions is small as reflected in fig. 6; the antenna azimuth angle is adjusted to be the RSRP value reported by the UE in the 150-degree domain, and the statistical histogram of the predicted value of the received power of the UE in the domain obtained by the algorithm is shown in figures 7 and 8, respectively, and the statistical rules of the RSRP value and the statistical histogram are basically consistent.
Example 2: the same as the specific calculation procedure in example 1;
according to the industrial parameter record, the maximum transmitting power of the base station is 53dBm (which is converted into linear power of about 200W), the channel number of the central carrier frequency is 37900, the antenna gain is 14dBi, 34459 user points are needed when the initial azimuth angle of the antenna is 130 degrees, if the azimuth angle is adjusted to 150 degrees, the field intensity distribution of the coverage area is required to be predicted, and 17849 user points are needed when the azimuth angle of the antenna is 150 degrees for comparison calculation of prediction errors. The prediction result shows that the method predicts the field intensity distribution after the base station antenna is adjusted, the average value of the absolute value of the error reaches 7.40dB, and the prediction error of a distribution area with more than 90% is less than 15dB; the prediction error distribution is shown in fig. 9, and the result in fig. 9 reflects that the prediction error of most regions is small;
the antenna azimuth angle is adjusted to be 150 degrees, the RSRP value reported by the UE in the domain and the statistical histogram of the predicted value of the received power of the UE in the domain obtained by the algorithm are respectively shown in figures 10 and 11, and the statistical rules of the RSRP value and the statistical histogram are basically consistent; in addition, the RSRP reported by the UE in the signal coverage area and the distribution of the UE received power predicted by the algorithm of the present invention are respectively shown in fig. 12 and fig. 13, and the two graphs intuitively reflect that the values and the distribution trend thereof are basically consistent;
obviously, it is an interesting problem that the received power of the UE at the same location is randomly changed due to the influence of many factors such as the mobile phone model, the mobile phone gesture, the position, the shielding object and other surrounding environments, and what probability distribution is satisfied. In this regard, the cell is divided into many square grids of 5 x 5m, and since the grid is already small enough (on the same order of magnitude as civilian GPS positioning accuracy), the received power of the grid may represent the received power at any point within the grid. We count the user data of 8 days when the antenna angle is unchanged for each grid of two embodiments, wherein the RSRP statistical histogram and the fitted positive-ethernet distribution curve reported by the UE in a certain grid when the antenna azimuth angle is unchanged for embodiment 1 are shown in fig. 14, and the RSRP statistical histogram and the fitted positive-ethernet distribution curve reported by the UE in a certain grid when the antenna azimuth angle is unchanged for embodiment 2 are shown in fig. 15.
In the figure, the horizontal axis represents power, the vertical axis represents the occurrence frequency of user data in each power segment, and the red line represents a normal distribution curve which is drawn by adopting the mean and the variance as predicted values of normal distribution parameters. In total 1158 pieces of user data are reported in a certain grid of the embodiment 1 shown in fig. 14, the average value of all pieces of user data of the grid is-103.72 dBm, the standard deviation is 4.60dBm, and the 99% confidence interval width is about 13.8dB; in total 1185 pieces of user data are reported in a certain grid of the embodiment 2 shown in fig. 15, the average value of all pieces of user data of the grid is-100.57 dBm, the standard deviation is 6.02dBm, and the 99% confidence interval width is about 18.06dB; as can be seen from fig. 14 and 15, the received power reported by the UEs in the same grid with the antenna azimuth angle unchanged is basically consistent with the statistical rule of the normal distribution, so the UE is considered to satisfy the quasi-normal distribution, and the width of the 99% confidence interval of each grid is generally about 15dB through the statistical analysis of all grid powers in two embodiments, so the error value of the prediction model is acceptable in engineering.
Fig. 16-17 show the variation of the statistical law of the received power of the same grid before and after the antenna adjustment, although the mean value is changed from-91.35 dBm to-98.72 dBm, the standard deviation is changed from 6.30dBm to 5.72dBm only, and is changed by 0.58dBm, so the standard deviation of the same grid before and after the antenna adjustment is regarded as being unchanged.
It will be apparent to those skilled in the art that the present invention has been described in detail by way of illustration only, and it is not intended to be limited by the above-described embodiments, as long as various insubstantial modifications of the method concepts and aspects of the invention are employed or the inventive concepts and aspects of the invention are directly applied to other applications without modification, all within the scope of the invention.
Claims (6)
1. The method for predicting the field intensity distribution of the cell after the antenna adjustment based on the parameters and the BP neural network is characterized by comprising the following steps:
(1) Collecting user MDT data, and reporting the user MDT data by User Equipment (UE);
(2) Processing the MDT data collected in the step (1) and removing abnormal values;
(3) Dividing a cell into square grids, dividing the divided square grids into two types, wherein the first type of grids are grids with sufficient RSRP data and are used for training a neural network; the second type of grids are grids with insufficient RSRP data;
(4) Training by adopting a three-layer neural network, and outputting grid average received power with insufficient RSRP data;
(5) And obtaining the average received power of each grid after the antenna azimuth rotation according to the formula of the average received power of each grid after the antenna azimuth rotation, namely, predicting the field intensity distribution of the cell after the antenna adjustment.
2. The method for predicting the field strength distribution of the cell after antenna adjustment based on the parameters and the BP neural network as set forth in claim 1, wherein in said step (1)The user MDT data of (1) includes longitude θ and latitude of the userTime advance T a And reference signal received strength RSRP.
3. The method for predicting the field strength distribution of the cell after antenna adjustment based on the parameters and the BP neural network according to claim 2, wherein,
the step (2) specifically comprises the following steps:
s21, taking the point with the altitude less than zero as abnormal data, and deleting the abnormal data;
s22, setting the altitude data as a constant 0, carrying out two-dimensional ink card support projection on the longitude and latitude of the reported user data, and enabling the north direction to correspond to the x-axis direction of the coordinates, wherein the azimuth angle of the antenna is the anticlockwise included angle between the vector from the antenna to the test point and the positive axis of the x; subtracting the antenna coordinates from the user point coordinates to obtain the relative coordinates of the user points with the antenna as an origin as new coordinates;
s23, dividing an antenna radiation coverage area into a plurality of grids, wherein the size of the grids is in the same order of magnitude as the precision of a Global Positioning System (GPS); for the RSRP value reported by different User Equipments (UE) in each grid, taking the RSRP values with the mean value mu and standard deviation sigma respectively set as the RSRP values of the User Equipments (UE) at the position as the mean value and standard deviation of random variables, whereinm is the total number of user equipments reporting RSRP values within one grid,and the average value mu and the standard deviation sigma of the data reported by the users in the grid are calculated, and the data positioned in [ mu-3 sigma, mu+3 sigma is obtained]Data within the range, data outside the range are used as outliers and deleted.
4. The method for predicting the field strength distribution of a cell after antenna adjustment based on parameters and BP neural network as set forth in claim 3,
the step (3) specifically comprises the following steps:
dividing a cell into square grids of n M multiplied by n M, replacing the coordinates of User Equipment (UE) positioned in a certain grid with the coordinates of the center of the grid, taking the received power of the User Equipment (UE) in the grid as a random variable, and calculating the mean value and standard deviation of the random variable based on RSRP data reported by all the User Equipment (UE) in the grid; dividing all grids into two types according to different RSRP data amounts, wherein the first type of grids are grids with sufficient RSRP data and are used for training a neural network; the second type of grids are grids with insufficient RSRP data and even unreported RSRP data, and the corresponding grid power is predicted through a neural network.
5. The method for predicting the field strength distribution of a cell after antenna adjustment based on parameters and BP neural network as recited in claim 4, wherein,
the step (4) specifically comprises the following steps:
training a grid with sufficient RSRP data by using a three-layer neural network, wherein the input layer is the coordinates (x a ,y a ) Wherein a represents a grid index number; the number of neurons contained in the middle hidden layer and the scale of a sample set adopted by training are in positive correlation, the output layer is a predicted value of the average value of the received power of the second type of grids, and constant term neurons of the input layer and the hidden layer belong to a bias unit;
wherein the activation function employed by each neuron of the hidden layer is:
wherein θ is hidden The weight vector from the input layer to the hidden layer neuron forms a weight matrix from the input layer to the hidden layer;
the activation function adopted by the output layer is a linear function:
where x is the vector of the output components of the hidden layer, θ output The weight vectors of the neurons of all the output layers form a weight matrix from the hidden layer to the output layer; different weight matrixes are arranged between layers of the neural network, the values of the weight matrixes are initialized randomly, a final weight matrix value is obtained along with training of the neural network, and finally, the predicted value of the average value of all grid received power in an effective radiation area of the antenna before the antenna is adjusted is obtained.
6. The method for predicting the field strength distribution of a cell after antenna adjustment based on the parameters and the BP neural network as recited in claim 5, wherein,
the step (5) specifically comprises the following steps:
obtaining an average received power formula of each grid after antenna azimuth rotation by adopting an electric wave propagation formula through an electric wave propagation model in a field intensity prediction model, and predicting user received power in an effective coverage area after antenna azimuth adjustment based on the electric wave propagation formula and antenna pattern gain change, namely obtaining prediction of cell field intensity distribution after antenna adjustment;
the electric wave propagation formula is as follows:
pt in this formula is the antenna transmit power,GT is antenna gain, lambda is wave wavelength, d is free space distance from antenna to user point, alpha is path loss index, pr is receiving power of receiving point;
the average received power per grid after antenna azimuth rotation is given by:
wherein the method comprises the steps ofAnd->Respectively representing the rotation front and back (azimuth angle, downtilt angle) of the antenna, comparing the rotation front and back (azimuth angle, downtilt angle) with the average value of the user measured data reported after the antenna posture adjustment, and taking the obtained error as the error of the average value of the receiving power predicted by the algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911263143.0A CN111062466B (en) | 2019-12-11 | 2019-12-11 | Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911263143.0A CN111062466B (en) | 2019-12-11 | 2019-12-11 | Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111062466A CN111062466A (en) | 2020-04-24 |
| CN111062466B true CN111062466B (en) | 2023-08-04 |
Family
ID=70300474
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911263143.0A Active CN111062466B (en) | 2019-12-11 | 2019-12-11 | Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111062466B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112013833B (en) * | 2020-07-23 | 2022-08-05 | 中国人民解放军战略支援部队信息工程大学 | VICTS antenna satellite-to-satellite tracking method based on deep neural network |
| CN114070676B (en) * | 2020-08-05 | 2023-03-14 | 展讯半导体(南京)有限公司 | Method and device for reporting and receiving AI network model support capability and storage medium |
| CN112070721B (en) * | 2020-08-13 | 2024-01-12 | 五邑大学 | Antenna parameter measurement method, device and storage medium based on instance division network |
| CN112187387A (en) * | 2020-09-22 | 2021-01-05 | 北京邮电大学 | Novel reinforcement learning method based on rasterization user position automatic antenna parameter adjustment |
| CN114599047A (en) * | 2020-12-03 | 2022-06-07 | 中兴通讯股份有限公司 | Network parameter derivation method, electronic device and storage medium |
| CN115913486A (en) * | 2021-09-18 | 2023-04-04 | 维沃移动通信有限公司 | Information reporting method, device, terminal and readable storage medium |
| CN116528273B (en) * | 2023-06-30 | 2023-10-31 | 中国电信股份有限公司 | Received power calculation method and device, storage medium and electronic equipment |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108307397A (en) * | 2017-01-13 | 2018-07-20 | 中国移动通信集团四川有限公司 | Network coverage evaluation method and system |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7774143B2 (en) * | 2002-04-25 | 2010-08-10 | The United States Of America As Represented By The Secretary, Department Of Health And Human Services | Methods for analyzing high dimensional data for classifying, diagnosing, prognosticating, and/or predicting diseases and other biological states |
-
2019
- 2019-12-11 CN CN201911263143.0A patent/CN111062466B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108307397A (en) * | 2017-01-13 | 2018-07-20 | 中国移动通信集团四川有限公司 | Network coverage evaluation method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111062466A (en) | 2020-04-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111062466B (en) | Method for predicting field intensity distribution of cell after antenna adjustment based on parameters and neural network | |
| Masood et al. | A machine learning based 3D propagation model for intelligent future cellular networks | |
| Ahmadien et al. | Predicting path loss distribution of an area from satellite images using deep learning | |
| Braham et al. | Fixed rank kriging for cellular coverage analysis | |
| CN111147163A (en) | Wireless communication link loss prediction method based on DNN neural network | |
| Butt et al. | RF fingerprinting and deep learning assisted UE positioning in 5G | |
| Moraitis et al. | Machine learning-based methods for path loss prediction in urban environment for LTE networks | |
| CN106886552A (en) | Location fingerprint database update method and system | |
| CN108064075B (en) | Method and apparatus for adjusting reference signal transmit power of one or more cells | |
| Isabona et al. | Hybrid neural network approach for predicting signal propagation loss in urban microcells | |
| Yin et al. | Proximity report triggering threshold optimization for network-based indoor positioning | |
| Qureshi et al. | Enhanced MDT-based performance estimation for AI driven optimization in future cellular networks | |
| Thrane et al. | Drive test minimization using deep learning with Bayesian approximation | |
| Varshney et al. | Quasi-deterministic channel propagation model for an urban environment at 28 GHz | |
| Fernandes et al. | Comparison of artificial intelligence and semi-empirical methodologies for estimation of coverage in mobile networks | |
| CN113873532A (en) | Intelligent park 5G network planning method | |
| Kim et al. | mmWave path loss modeling for urban scenarios based on 3D-convolutional neural networks | |
| Fernández Anitzine et al. | Influence of training set selection in artificial neural network-based propagation path loss predictions | |
| Sotiroudis et al. | Optimal artificial neural network design for propagation path-loss prediction using adaptive evolutionary algorithms | |
| Wang et al. | Mobile device localization in 5G wireless networks | |
| Nguyen et al. | Deep Learning for Path Loss Prediction at 7 GHz in Urban Environment | |
| Bärring et al. | Factory radio design of a 5G network in offline mode | |
| Yi et al. | Neural network based prediction and analysis for NB-IoT network location | |
| Seretis et al. | Ray-tracing driven ann propagation models for indoor environments at 28 GHz | |
| Combeau et al. | A numerical simulation system for mobile telephony base station EMF exposure using smartphones as probes and a genetic algorithm to improve accuracy |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |